Wiki Log

Chronological record of all wiki actions. Append-only. Format: ## [YYYY-MM-DD] action | subject Actions: ingest, update, query, lint, create, archive, delete When this file exceeds 500 entries, rotate: rename to log-YYYY.md, start fresh.

[2026-05-01] create | Wiki initialized

• Domain: Personal knowledge spanning AI, investing, tech, productivity (multi-domain pattern)
• Path: /root/wiki
• Structure: SCHEMA.md, index.md, log.md, raw/{transcripts,articles,papers,assets}, entities/{ai,investing,tech,productivity}, concepts/{ai,investing,tech,productivity}, comparisons, queries
• Taxonomy: domain/* namespace (one per page) + type-specific tags

[2026-05-01] ingest | Jensen Huang × Cleo Abram — Huge Conversations (2026)

• Source: https://youtu.be/7ARBJQn6QkM (~1h02m, 58638 bytes)
• Raw: raw/transcripts/jensen-huang-cleo-abram-2026.md (sha256 2d83b5d1c64a...)
• Created:
• entities/ai/jensen-huang.md
• entities/ai/nvidia.md
• entities/tech/cleo-abram.md
• concepts/ai/physical-ai.md
• queries/jensen-huang-cleo-abram-2026.md (source-index)
• Updated: index.md (5 new entries)
• Cross-links: 11 wikilinks total across the 5 new pages

2026-05-01 — Ingest: Jensen × Lex Fridman (2026)

• source: https://youtu.be/vif8NQcjVf0
• transcript: raw/transcripts/jensen-huang-lex-fridman-2026.md (sha256 85b3c787b0bb…, 145,447 bytes, 2h25m)
• new entities: lex-fridman
• new concepts: extreme-co-design, install-base-moat, scaling-laws-plural, ai-supply-chain
• updated: jensen-huang, nvidia, physical-ai (added Lex source + new cross-links)
• new query: jensen-huang-lex-fridman-2026

[2026-05-01] ingest | Andrej Karpathy × Sequoia AI Ascent (2026)

• Source: https://youtu.be/96jN2OCOfLs (~30min, 33,232 plain chars / 40,376 ts chars)
• Raw: raw/transcripts/andrej-karpathy-sequoia-2026.md (+ .ts.txt, .plain.txt; sha256 dcc52687...)
• New entities: andrej-karpathy, sequoia-capital
• New concepts: vibe-coding, agentic-engineering, software-3-0, jagged-intelligence, animals-vs-ghosts, llm-knowledge-bases
• New query: andrej-karpathy-sequoia-2026
• Updated: index.md (9 new entries, total pages 11 → 20)
• Cross-links: ~30 wikilinks total across the 9 new pages

[2026-05-01] ingest | Mitchell Hashimoto × Zed — Agentic Engineering in Action (2025-06-18)

• Source: https://youtu.be/XyQ4ZTS5dGw (~1h01m, 70,644 plain chars / ~76k ts chars)
• Raw: raw/transcripts/mitchell-hashimoto-zed-agentic-2025.md (+ .ts.txt, .plain.txt; sha256 a380d0f9...)
• New entities: mitchell-hashimoto, ghostty, zed-editor, claude-code, jujutsu
• New concepts: agent-as-junior-engineer, driving-into-mud, parallel-agent-competitions
• New query: mitchell-hashimoto-zed-agentic-2025
• Updated: index.md (9 new entries, total pages 20 → 29)
• Cross-links: ~35 wikilinks total across the 9 new pages
• Bridges to existing agentic-engineering (Karpathy's frame, reinforced), vibe-coding, jagged-intelligence

2026-05-01 — Ingest: Eric Zakariasson × AI Engineer 2026 (software factory)

• source_url: https://youtu.be/rnDm57Py54A
• video_id: rnDm57Py54A
• duration: 5003s (~83min), 72805 plain chars, 2019 segments
• sha256: fb593a7ce47a8dd851b689e6d95c57511c0ca5ecdfb2c242955d00a16e7bf51a
• raw: raw/transcripts/eric-zakariasson-software-factory-2026.{md,ts.txt,plain.txt}
• new entities: eric-zakariasson, cursor
• new concepts: software-factory, levels-of-autonomy-shapiro, emergent-cursor-rules, verifiable-systems-for-agents, isolated-agent-vms, sync-plan-async-execute
• new query: eric-zakariasson-software-factory-2026
• pages: 29 → 39 (+10)
• cross-links: andrej-karpathy, vibe-coding, agentic-engineering, mitchell-hashimoto, agent-as-junior-engineer, parallel-agent-competitions, driving-into-mud, jagged-intelligence, claude-code, zed-editor

2026-05-01 — Ingest: Peter Steinberger × AI Engineer 2026 (State of the Claw)

• source_url: https://youtu.be/zgNvts_2TUE
• video_id: zgNvts_2TUE
• duration: 2640s (~44min), 37205 plain chars
• sha256: 9e463e839c579752f5203888242dd0fda4b473510d1594f5d6990ca4c81bf095
• raw: raw/transcripts/peter-steinberger-state-of-the-claw-2026.{md,ts.txt,plain.txt}
• new entities: peter-steinberger, openclaw, openclaw-foundation
• new concepts: agent-taste, agent-security-slop, soul-md
• new query: peter-steinberger-state-of-the-claw-2026
• pages: 39 → 47 (+8)
• cross-links: andrej-karpathy, eric-zakariasson, mitchell-hashimoto, ghostty, vibe-coding, agentic-engineering, llm-knowledge-bases, animals-vs-ghosts, driving-into-mud, emergent-cursor-rules

2026-05-01 — Ingest: Mahmoud Mabrouk × AI Engineer 2026 (Judge the Judge / GEPA)

• source_url: https://youtu.be/X4dEHRzBLmc
• video_id: X4dEHRzBLmc
• duration: 2452s (~41min), 34769 plain chars
• sha256: a715d17320bc24c3252ae4c784f93279321bf50c2a00290d8c54e1338f685ee6
• raw: raw/transcripts/mahmoud-mabrouk-judge-the-judge-gepa-2026.{md,ts.txt,plain.txt}
• new entities: mahmoud-mabrouk, agenta
• new concepts: llm-as-a-judge, llm-judge-calibration, gepa
• new query: mahmoud-mabrouk-judge-the-judge-gepa-2026
• pages: 47 → 54 (+7)
• cross-links: verifiable-systems-for-agents, jagged-intelligence, dspy, emergent-cursor-rules

2026-05-01 — Ingest: Harshil Agrawal × AI Engineer 2026 (Sandbox AI-Generated Code)

• source_url: https://youtu.be/AHtGAgQ0Q_Q
• video_id: AHtGAgQ0Q_Q
• duration: 2306s (~38min), 28924 plain chars
• raw: raw/transcripts/harshil-agrawal-sandbox-ai-code-2026.{md,ts.txt,plain.txt}
• new entities: harshil-agrawal, cloudflare
• new concepts: ai-generated-code-is-untrusted, capability-based-security, isolates-vs-containers
• new query: harshil-agrawal-sandbox-ai-code-2026
• pages: 54 → 61 (+7)
• cross-links: agent-security-slop, verifiable-systems-for-agents, isolated-agent-vms, parallel-agent-competitions, animals-vs-ghosts

2026-05-01 — Ingest: Patrick Riley & Carlos Galan × AI Engineer 2026 (Identity for AI Agents — Auth0)

• source_url: https://youtu.be/VSdV-AdSlis
• video_id: VSdV-AdSlis
• duration: 4932s (~82min), 62182 plain chars
• raw: raw/transcripts/riley-galan-identity-for-ai-agents-auth0-2026.{md,ts.txt,plain.txt}
• new entities: patrick-riley, carlos-galan, auth0
• new concepts: auth-for-ai-four-pillars, token-vault, async-auth-ciba, agent-as-oauth-client, dynamic-client-registration
• new query: riley-galan-identity-for-ai-agents-auth0-2026
• pages: 61 to 71 (+10)
• cross-links: capability-based-security, ai-generated-code-is-untrusted, isolates-vs-containers, soul-md, isolated-agent-vms, verifiable-systems-for-agents, agent-security-slop, llm-judge-calibration

2026-05-01 — Batch summary

Ingested 4 AI Engineer 2026 sessions in one session (Steinberger, Mabrouk, Agrawal, Riley+Galan) forming a trust-boundary quadrilateral: supply-side noise (Steinberger) x eval-side theater (Mabrouk) x execution-side sandboxing (Agrawal) x identity-side delegation (Riley+Galan). See queries for cross-ingest synthesis.

2026-05-01 — Ingest: Ryan Lopopolo × AI Engineer London 2026 (Harness Engineering)

• source_url: https://youtu.be/am_oeAoUhew
• video_id: am_oeAoUhew
• duration: 2775s (~46min), 43135 plain chars
• raw: raw/transcripts/ryan-lopopolo-harness-engineering-2026.{md,ts.txt,plain.txt}
• new entities: ryan-lopopolo, openai
• new concepts: harness-engineering, code-is-free, non-functional-requirements-as-prompts
• new query: ryan-lopopolo-harness-engineering-2026
• pages: 71 → 76 (+5)
• cross-links: agentic-engineering, software-factory, agent-as-junior-engineer, driving-into-mud, emergent-cursor-rules, parallel-agent-competitions, llm-judge-calibration, agent-taste, verifiable-systems-for-agents, soul-md

2026-05-01 — Ingest: Maggie Konstanty (MLOps.community #372)

• Source: https://youtu.be/9EjWR3QpJYk · 2457s · sha 8ee75034f1f31cb6 · 39789 chars
• New entities: maggie-konstanty, prosus, arize
• New concepts: eval-lifecycle-pre-to-production, silent-failure-dropoff, error-analysis-as-detective-work
• New query: maggie-konstanty-evals-2026
• Cross-links: llm-as-a-judge, llm-judge-calibration, gepa, agent-taste, harness-engineering
• Wiki count: 76 → 82

2026-05-01 · Bulk ingest · AI Engineer London 2026 + Coding Agents Conference 2026 + MLOps Podcast #370

Ingested 8 transcripts in a single batch (user: "Finish it all"):

1. Faye Zhang (Pinterest) — Production Sub-agents for LLM Post-Training · https://youtu.be/kPC4YOkIxVo
2. Entities: faye-zhang, pinterest
3. Concepts: subagent-architecture, agentic-posttraining-failure-modes, hot-celebrity-orchestration-problem

4. Query: faye-zhang-subagents-posttraining-2026

5. Erin Ahmed (Cleric) — Self-Learning Agent Memory · https://youtu.be/tKRel4kEpSI

6. Entities: erin-ahmed, cleric
7. Concepts: learning-agent-loop, correction-must-persist-compound-visible, ambient-vs-directed-learning

8. Query: erin-ahmed-learning-agent-2026

9. Ash Lewis (Fastino) — Choosing the Right Model / Pioneer · https://youtu.be/HVqyWXa1SJ4

10. Entities: ash-lewis, fastino
11. Concepts: model-drift, agents-as-model-selectors, inference-logs-as-training-data

12. Query: ash-lewis-fastino-model-drift-2026

13. Zach Lloyd (Warp) — 2026 Year of Agent Orchestration · https://youtu.be/eT1F2BAZJ64

14. Entities: zach-lloyd, warp
15. Concepts: agent-orchestration-2026, cloud-agent-primitives, laptop-limit

16. Query: zach-lloyd-agent-orchestration-2026

17. Niels Bantilan (Union.ai / Flyte) — Decomposing the Agent Orchestration System · https://youtu.be/H9fsxdK-NeQ

18. Entities: niels-bantilan, union-ai, flyte
19. Concepts: durable-observable-debuggable-agents, replay-log, infrastructure-as-context

20. Query: niels-bantilan-orchestration-lessons-2026

21. Harrison Chase + Sam Partee (LangChain + Arcade) — Coding Agent → General Purpose Agent · https://youtu.be/OYK1-vCbSMU

22. Entities: harrison-chase, sam-partee, langchain, arcade-ai
23. Concepts: coding-agent-as-foundation, tool-runtime-vs-harness, delegated-agent-authorization

24. Query: harrison-sam-general-purpose-agents-2026

25. Dexter Horthy (HumanLayer) — Everything We Got Wrong About Research-Plan-Implement · https://youtu.be/YwZR6tc7qYg

26. Entities: dexter-horthy, humanlayer
27. Concepts: qrspi, do-not-outsource-thinking, dumb-zone, static-asset-workflow, control-flow-vs-prompt-flow, context-engineering

28. Query: dexter-horthy-qrspi-2026

29. Mihail Eric (Monaco / Stanford CS146S) — The Modern Software Developer · https://youtu.be/jOe4fJSc2IE

30. Entities: mihail-eric, monaco, modern-software-developer-course
31. Concepts: engineer-as-director-of-agents, upfront-investment-paradox, verifiability-frontier
32. Query: mihail-eric-modern-software-engineer-2026

Wiki now at 134 pages (48 entities, 67 concepts, 19 queries).

[2026-05-01] ingest | Sanchez · DevOps Playbook for the Agentic Era (Microsoft DevBlogs 2026)

• Source: raw/articles/david-sanchez-agentic-devops-2026.md (sha 4b4b758acd9062e7)
• New entities: david-sanchez, microsoft, github-copilot
• New concepts: agents-scale-not-fix, repository-as-interface, four-collaboration-zones, specification-driven-development, pipeline-as-verifier, pipeline-as-specification, agentic-devops-maturity-model
• New query: david-sanchez-agentic-devops-2026
• Back-links added to: engineer-as-director-of-agents, do-not-outsource-thinking, qrspi, upfront-investment-paradox, harness-engineering, delegated-agent-authorization, coding-agent-as-foundation, ai-generated-code-is-untrusted, verifiability-frontier, verifiable-systems-for-agents, durable-observable-debuggable-agents
• Wiki now: 145 pages (51 entities, 74 concepts, 20 queries)

[2026-05-02] ingest | Parakhin · AI-Native Engineering (Latent Space 2026)

• Source: raw/transcripts/mikhail-parakhin-latent-space-2026.md (sha 92a2eb5ebbe6b780)
• New entities: mikhail-parakhin, shopify, liquid-ai, swyx, latent-space
• New concepts: shopify-ai-adoption-inflection, tokens-need-critique-loop, auto-research-loop, tangle, tangent, simgym, simulation-data-moat, hstu, ucp-catalog, sidekick-pulse, distill-to-small-task-model
• New query: mikhail-parakhin-shopify-ai-stack-2026
• Cross-links into: install-base-moat, ai-generated-code-is-untrusted, agents-scale-not-fix, do-not-outsource-thinking, coding-agent-as-foundation, animals-vs-ghosts, pipeline-as-verifier, pipeline-as-specification, durable-observable-debuggable-agents, non-functional-requirements-as-prompts, llm-as-a-judge, llm-judge-calibration, scaling-laws-plural
• Wiki now: 183 pages (56 entities, 85 concepts, 21 queries, 20 transcripts, 1 article)

2026-05-02 — Ryan Lopopolo on Latent Space (Symphony / Extreme Harness Engineering)

• Source: https://youtu.be/CeOXx-XTYek · Latent Space · ~1h17m · sha e844182f7bb53419
• Raw: ryan-lopopolo-latent-space-symphony-2026
• Query: ryan-lopopolo-latent-space-symphony-2026 (query page with same slug under queries/)
• New entities: openai-frontier
• New concepts: symphony, ghost-library, agent-legible-software, skill-distillation, dual-agent-pr-loop
• Updated: ryan-lopopolo (added Frontier affiliation + 2nd source), harness-engineering (Reifications section), code-is-free (ghost-library xref)
• Index totals: 191 pages (entities=59, concepts=91, queries=22, transcripts=21, articles=1)
• Companion to prior ingest ryan-lopopolo-harness-engineering-2026 — this interview extends the conference talk with Symphony internals, the 0%-human-reviewed PR loop, and OpenAI Frontier's enterprise strategy.

2026-05-01 — Ghost package: @steipete/summarize/prompts

• Target: steipete/summarize @ 7a14750 (v0.14.1)
• Scope: packages/core/src/prompts/* + minimal deps (shared/contracts.ts, language.ts) — pure string-building only
• Location: /root/wiki/ghosts/summarize/
• Operations covered: 14 (6 constants + 4 helpers + 4 builders)
• Evidence: 61 observations, SHA-256 1989a5d4…
• Parity: 2-run SHA match ✓
• Mutations: 4/4 detected (threshold swap, wording tweak, token-count change, envelope reorder)
• Key decision: scoped to prompts only; content extraction / transcript / CLI / extension excluded (stateful, poor ghost fit)
• Cross-link: ghost-library, summarize-sh-prompt (existing skill uses these same prompt builders)

[2026-05-02] ingest | Chris Shayan (Backbase) — Intelligence Layer for Banking Agents

• Source: LinkedIn post by Chris Shayan, Head of AI, Backbase
• Files created:
• raw/articles/chris-shayan-backbase-intelligence-layer-2026.md
• entities/ai/chris-shayan.md
• entities/ai/backbase.md
• queries/chris-shayan-backbase-intelligence-layer-2026.md
• Files updated:
• index.md (+2 entities, +1 query, +1 raw article; totals bumped to 194)
• Framing: first regulated-industry (banking) practitioner in wiki to independently converge on the harness-centric architecture. Shayan's Signal Catalogue / Digital Twin / Nudge Mesh map 1:1 to control-flow-vs-prompt-flow / learning-agent-loop / silent-failure-dropoff + levels-of-autonomy-shapiro.
• Flagged tensions: regulation-capped variant of harness thesis (L5/L6 may be structurally inaccessible in banking); "silence as an action" candidate for new concept page.

2026-05-02 — Ingest: Florian Juengermann × LangChain Max Agency (Listen agents)

• source_url: https://youtu.be/YTTH-0XXEBE
• video_id: YTTH-0XXEBE
• duration: 2858s (~47min), 58284 plain chars
• sha256: 8f199bbda2f19136...
• raw: raw/transcripts/florian-juengermann-listen-agents-2026.{md,ts.txt,plain.txt}
• new entities: florian-juengermann, listen-labs
• new concepts: virtual-table-architecture, map-reduce-classification, feedback-subagent, contextual-prompt-engineering, powerpoint-subagent, live-report-numbers, trigger-long-running-agent, composer-agent-ux
• new query: florian-juengermann-listen-agents-2026
• index regenerated from filesystem walk (11 new pages)
• cross-links: harrison-chase, langchain, claude-code, subagent-architecture, skill-distillation, llm-as-a-judge, eval-lifecycle-pre-to-production, context-engineering, harness-engineering, ryan-lopopolo, mitchell-hashimoto, faye-zhang-subagents-posttraining-2026, harrison-sam-general-purpose-agents-2026, mahmoud-mabrouk-judge-the-judge-gepa-2026, maggie-konstanty-evals-2026, software-3-0, mihail-eric-modern-software-engineer-2026, model-drift, sync-plan-async-execute, agent-as-junior-engineer, agent-taste, silent-failure-dropoff, emergent-cursor-rules, isolates-vs-containers, distill-to-small-task-model, soul-md, peter-steinberger-state-of-the-claw-2026, dexter-horthy-qrspi-2026

2026-05-02 — Ingest: Jared Zoneraich × AI Engineer "How Claude Code Works"

• source_url: https://youtu.be/RFKCzGlAU6Q
• video_id: RFKCzGlAU6Q
• duration: ~1h05m, 61839 plain chars
• sha256: 8faf0119a4a2c4751ad831bb26974ccba34d3ced96f3092015cbd1c63bbf0927
• raw: raw/transcripts/jared-zoneraich-claude-code-works-2026.{md,ts.txt,plain.txt}
• new entities: jared-zoneraich, promptlayer
• new concepts: claude-code-master-loop, claude-code-todo-tool, context-compaction, bash-as-universal-tool, claude-code-skills, agent-smell
• updated: claude-code (+Architecture section, reverse-engineered by Zoneraich — flagged as unofficial, confidence medium; bumped updated, added second source)
• new query: jared-zoneraich-claude-code-works-2026
• cross-links: mitchell-hashimoto-zed-agentic-2025, peter-steinberger-state-of-the-claw-2026, florian-juengermann-listen-agents-2026, ryan-lopopolo-harness-engineering-2026, harness-engineering, context-engineering, subagent-architecture, agentic-engineering, contextual-prompt-engineering, skill-distillation, eval-lifecycle-pre-to-production
• Framing: the missing architectural lower-layer beneath the wiki's existing Claude Code coverage. Hashimoto gave usage patterns; Steinberger gave ecosystem view; Zoneraich (unofficial) gives the internals — master loop nO, H2A compaction at ~92%, todo tool as prompt-enforced plan, task tool as sub-agent spawn, skills as extendable system prompts, bash as the load-bearing universal tool.
• index.md NOT modified per task instruction (parent will regenerate).

2026-05-02 — Ingest: Alex Krentsel — Principles for Autonomous System Design (OpenClaw Deep Dive)

• source_url: https://youtu.be/sxX8BMscce0
• video_id: sxX8BMscce0
• duration: ~1h03m, 59581 plain chars
• sha256: 979c75f11571...
• raw: raw/transcripts/alex-krentsel-openclaw-deep-dive-2026.{md,ts.txt,plain.txt}
• new entities: alex-krentsel
• new concepts: phase-3-autonomous-agents, agentic-loopiness, sessions-as-processes, heartbeat-and-cron, skills-over-mcp, design-over-implementation
• new query: alex-krentsel-openclaw-deep-dive-2026
• updated (2nd source added, architecture detail appended): openclaw (major — full three-layer architecture section, gateway controller internals, ACP, Ludwig case studies), peter-steinberger, openclaw-foundation, agent-taste, soul-md, agent-security-slop
• companion to peter-steinberger-state-of-the-claw-2026 — Peter = culture/security angle; Krentsel = systems/architecture angle (sessions-as-processes, heartbeat+cron, skills>MCP, design>implementation)
• total new content: ~24k chars across 8 new files; ~5k chars appended to 6 existing
• index regen deferred to parent

2026-05-02 — Ingest: Matt Pocock — Full Walkthrough: Workflow for AI Coding (AI Engineer 2026)

• source_url: https://youtu.be/-QFHIoCo-Ko
• video_id: -QFHIoCo-Ko
• duration: 1h36m25s, 89753 plain chars
• sha256: ea6f5916e5566898aacf0a72cc78ee21e48c21de33be38f558923227d48651a6
• raw: raw/transcripts/matt-pocock-ai-coding-workflow-2026.{md,ts.txt,plain.txt}
• FILTERING: user-curated ingest. 2-hour workshop with substantial filler (breaks, audience surveys, jokes, live-coding QA); only the 8 themes in the user's scaffold were distilled. Breaks/surveys/jokes NOT captured per explicit user direction.
• new entities: matt-pocock
• new concepts: smart-vs-dumb-zones, memento-principle, grill-me-skill, prd-as-destination, tracer-bullets, afk-implementation, tdd-feedback-loop, push-vs-pull-standards
• new query: matt-pocock-ai-coding-workflow-2026
• cross-links: claude-code, claude-code-skills, claude-code-master-loop, claude-code-todo-tool, context-compaction, context-engineering, contextual-prompt-engineering, skill-distillation, skills-over-mcp, feedback-subagent, harness-engineering, agent-legible-software, subagent-architecture, sessions-as-processes, agent-taste, agent-smell, agent-orchestration-2026, jared-zoneraich-claude-code-works-2026, florian-juengermann-listen-agents-2026, alex-krentsel-openclaw-deep-dive-2026, peter-steinberger-state-of-the-claw-2026, mitchell-hashimoto-zed-agentic-2025, ryan-lopopolo-harness-engineering-2026
• Notable tension flagged: memento-principle ↔ context-compaction — Pocock prefers /clear, Zoneraich's reverse-engineering showed H2A auto-compaction as load-bearing inside Claude Code's main loop. Both framings can coexist (externalised plan state vs autonomous long-run); marked via contradictions: frontmatter on memento-principle.
• skipped themes: none — all 8 user-scaffold themes minted.
• index.md NOT modified per task instruction (parent will regenerate).

2026-05-02 — Ingest: Stefano Fiorucci — Let LLMs Wander: Engineering RL Environments (AI Engineer 2026)

• source_url: https://youtu.be/71V3fTaUp2Q
• video_id: 71V3fTaUp2Q
• duration: ~40m23s, 29176 plain chars
• sha256: c81c05fb45be...
• raw: raw/transcripts/stefano-fiorucci-rl-environments-2026.{md,ts.txt,plain.txt}
• new entities: stefano-fiorucci
• new concepts: rl-environment-engineering, rl-with-verifiable-rewards, llm-wandering, verifiers-library, rl-curriculum-opponent-skill, synthetic-sft-bootstrap
• new query: stefano-fiorucci-rl-environments-2026
• cross-links: harness-engineering, eval-lifecycle-pre-to-production, verifiability-frontier, verifiable-systems-for-agents, faye-zhang-subagents-posttraining-2026, distill-to-small-task-model, skill-distillation, scaling-laws-plural, llm-as-a-judge, subagent-architecture, agentic-engineering, agent-as-junior-engineer
• framing: RL training counterpart to the inference/agent engineering talks already in the wiki. Fiorucci covers the DeepSeek R1 / O1 paradigm (RL with verifiable rewards), introduces the Verifiers library (Prime Intellect), and demonstrates a full SFT→CISPO RL pipeline on a small model (LFM-2 → tic-tac-toe master, outperforms teacher GPT-4o mini). Key lessons: batch size stability, hidden opponent bias, temperature as exploration dial, synthetic SFT bootstrap.
• index.md NOT modified per task instruction (parent will regenerate).

[2026-05-02] ingest | The Pragmatic Engineer — Mario Zechner & Armin Ronacher on Pi (self-modifying coding agent)

• raw/transcripts/mario-armin-pragmatic-pi-2026.md (new, sha256 4b58b2d6…)
• entities/ai/mario-zechner.md (new)
• entities/ai/armin-ronacher.md (new)
• entities/ai/pragmatic-engineer.md (new)
• entities/ai/pi-agent.md (new)
• concepts/ai/slow-the-f-down.md (new)
• concepts/ai/self-modifying-agent.md (new)
• concepts/ai/complexity-as-enemy.md (new)
• concepts/ai/case-against-mcp.md (new)
• concepts/ai/prompt-request-over-pull-request.md (new, contested)
• concepts/ai/dark-factory-skepticism.md (new, contested)
• concepts/ai/judgment-as-the-new-moat.md (new)
• queries/mario-armin-pragmatic-pi-2026.md (new)
• Cross-links into existing: peter-steinberger, openclaw, skills-over-mcp, harness-engineering, agents-scale-not-fix, do-not-outsource-thinking, engineer-as-director-of-agents, mitchell-hashimoto, ryan-lopopolo, design-over-implementation, code-is-free, agent-taste, agent-smell.